CLIP-blind pairs

[mahtabbigverdi]

Hi,
Great work, have you also released the code for finding clip blind pairs?

[clemneo]

+1, or even if the code isn't released it would be good if you could share some implementation details!

Are you taking the cosine similarity of the activations at the [CLS] token position at the last layer?

[tsb0601]

Hi, Sorry for the late reply. Of couse! We use the cosine similarity of the CLIPImage Encoder output: e.g. https://github.com/mlfoundations/open_clip/blob/58e4e39aaabc6040839b0d2a7e8bf20979e4558a/src/open_clip/model.py#L287

[namin-an]

How about DINO? @tsb0601 Thank you.